Menu

SphinxTrain:Creation of decision trees?

Help
shibli
2004-09-01
2012-09-22
  • shibli

    shibli - 2004-09-01

    I have been following the following steps:
    1) vector quantization
    2) making flat initial models (ci)
    3) initialization of flat models (ci)
    4) creation of cd_semi_untied models
    5) make_questions and then unpruned trees

    at this 5th step nothng happens in the trees directory only subdirectories are created with no contents. The log shows that calloc has failed and the warning that no triphones are defined for <PHONE>(any phone defined in my phone list).
    I have five utterances containing 'ONE', 'TWO', 'THREE','FOUR','FIVE'.

    Can someone elaborate on this.. i have tried all measures that i could think of.

     
    • The Grand Janitor

      Hi,
           If you used only 5 utterances to train the model, there is not even enough data for training the CI-models. 
      The script will not be suitable for you in that case. 
           Modern days acoustic model training using HMM use subwords models (like phoneme), each subword models usuallying require at least 100-200 samples to start.
           My suggestion is that you should start from existing corpus such as TIDIGITS or TIWORDS in your project.  Some of these database can be found in the web page of Linguistic Data Consortium and can be obtained in a very low price.  CMU also open source a couple of database such as the Communicator.
           If you want to collect data, you need to understand that this is a very time-consuming, you need to put it on your task list.
      Arthur

       
    • shibli

      shibli - 2004-09-02

      Thankyou very much, for clearing the matter.
      My task at this moment is to make Sphinx recognize the digits(0-9) and the words (YES and NO), do you think that i should use training or sould I use the model provided with Sphinx 3.

      Regards
      BILAL AHMED

       
    • The Grand Janitor

      Sphinx 3's default models is for large vocabulary.  In many cases, it proves to be too "flat", (or not a good match ) for the digit models.  I will suggest you to use thinks like TIDIGITS to do trainining first and collect some extra yes and no data yourself.  This is a situation your friends can help you a lot . :-)

       
    • shibli

      shibli - 2004-09-03

      well i acted on your advice ansd recorded some utterings which were taken from novels, now i have a corpus of about 100 utterances (raw format) with a length range of 4.5sec-12sec, making up of about 10min. 912 words, 40 phones even after so much work and data the build tree utilities (4th step) still fails giving me the same error FATAL ERROR calloc failed.
      What is the cause this time. and how can i resolve it. (THE TIDIGITS database costs $1900 which is too much (dont u think)) :)

      waiting for ur reply, with regards
      BILAL AHMED.

       

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.